Line data Source code
1 : use crate::pageserver_client::PageserverClient;
2 : use crate::persistence::Persistence;
3 : use crate::{compute_hook, service};
4 : use pageserver_api::controller_api::{AvailabilityZone, MigrationConfig, PlacementPolicy};
5 : use pageserver_api::models::{
6 : LocationConfig, LocationConfigMode, LocationConfigSecondary, TenantConfig, TenantWaitLsnRequest,
7 : };
8 : use pageserver_api::shard::{ShardIdentity, TenantShardId};
9 : use pageserver_client::mgmt_api;
10 : use reqwest::StatusCode;
11 : use std::borrow::Cow;
12 : use std::collections::HashMap;
13 : use std::sync::Arc;
14 : use std::time::{Duration, Instant};
15 : use tokio_util::sync::CancellationToken;
16 : use utils::backoff::exponential_backoff;
17 : use utils::generation::Generation;
18 : use utils::id::{NodeId, TimelineId};
19 : use utils::lsn::Lsn;
20 : use utils::pausable_failpoint;
21 : use utils::sync::gate::GateGuard;
22 :
23 : use crate::compute_hook::{ComputeHook, NotifyError};
24 : use crate::node::Node;
25 : use crate::tenant_shard::{IntentState, ObservedState, ObservedStateDelta, ObservedStateLocation};
26 :
27 : const DEFAULT_HEATMAP_PERIOD: &str = "60s";
28 :
29 : /// Object with the lifetime of the background reconcile task that is created
30 : /// for tenants which have a difference between their intent and observed states.
31 : pub(super) struct Reconciler {
32 : /// See [`crate::tenant_shard::TenantShard`] for the meanings of these fields: they are a snapshot
33 : /// of a tenant's state from when we spawned a reconcile task.
34 : pub(super) tenant_shard_id: TenantShardId,
35 : pub(crate) shard: ShardIdentity,
36 : pub(crate) placement_policy: PlacementPolicy,
37 : pub(crate) generation: Option<Generation>,
38 : pub(crate) intent: TargetState,
39 :
40 : /// Nodes not referenced by [`Self::intent`], from which we should try
41 : /// to detach this tenant shard.
42 : pub(crate) detach: Vec<Node>,
43 :
44 : /// Configuration specific to this reconciler
45 : pub(crate) reconciler_config: ReconcilerConfig,
46 :
47 : pub(crate) config: TenantConfig,
48 : pub(crate) preferred_az: Option<AvailabilityZone>,
49 :
50 : /// Observed state from the point of view of the reconciler.
51 : /// This gets updated as the reconciliation makes progress.
52 : pub(crate) observed: ObservedState,
53 :
54 : /// Snapshot of the observed state at the point when the reconciler
55 : /// was spawned.
56 : pub(crate) original_observed: ObservedState,
57 :
58 : pub(crate) service_config: service::Config,
59 :
60 : /// A hook to notify the running postgres instances when we change the location
61 : /// of a tenant. Use this via [`Self::compute_notify`] to update our failure flag
62 : /// and guarantee eventual retries.
63 : pub(crate) compute_hook: Arc<ComputeHook>,
64 :
65 : /// To avoid stalling if the cloud control plane is unavailable, we may proceed
66 : /// past failures in [`ComputeHook::notify`], but we _must_ remember that we failed
67 : /// so that we can set [`crate::tenant_shard::TenantShard::pending_compute_notification`] to ensure a later retry.
68 : pub(crate) compute_notify_failure: bool,
69 :
70 : /// Reconciler is responsible for keeping alive semaphore units that limit concurrency on how many
71 : /// we will spawn.
72 : pub(crate) _resource_units: ReconcileUnits,
73 :
74 : /// A means to abort background reconciliation: it is essential to
75 : /// call this when something changes in the original TenantShard that
76 : /// will make this reconciliation impossible or unnecessary, for
77 : /// example when a pageserver node goes offline, or the PlacementPolicy for
78 : /// the tenant is changed.
79 : pub(crate) cancel: CancellationToken,
80 :
81 : /// Reconcilers are registered with a Gate so that during a graceful shutdown we
82 : /// can wait for all the reconcilers to respond to their cancellation tokens.
83 : pub(crate) _gate_guard: GateGuard,
84 :
85 : /// Access to persistent storage for updating generation numbers
86 : pub(crate) persistence: Arc<Persistence>,
87 : }
88 :
89 : pub(crate) struct ReconcilerConfigBuilder {
90 : config: ReconcilerConfig,
91 : }
92 :
93 : impl ReconcilerConfigBuilder {
94 : /// Priority is special: you must pick one thoughtfully, do not just use 'normal' as the default
95 0 : pub(crate) fn new(priority: ReconcilerPriority) -> Self {
96 0 : Self {
97 0 : config: ReconcilerConfig::new(priority),
98 0 : }
99 0 : }
100 :
101 0 : pub(crate) fn secondary_warmup_timeout(self, value: Duration) -> Self {
102 0 : Self {
103 0 : config: ReconcilerConfig {
104 0 : secondary_warmup_timeout: Some(value),
105 0 : ..self.config
106 0 : },
107 0 : }
108 0 : }
109 :
110 0 : pub(crate) fn secondary_download_request_timeout(self, value: Duration) -> Self {
111 0 : Self {
112 0 : config: ReconcilerConfig {
113 0 : secondary_download_request_timeout: Some(value),
114 0 : ..self.config
115 0 : },
116 0 : }
117 0 : }
118 :
119 0 : pub(crate) fn tenant_creation_hint(self, hint: bool) -> Self {
120 0 : Self {
121 0 : config: ReconcilerConfig {
122 0 : tenant_creation_hint: hint,
123 0 : ..self.config
124 0 : },
125 0 : }
126 0 : }
127 :
128 0 : pub(crate) fn build(self) -> ReconcilerConfig {
129 0 : self.config
130 0 : }
131 : }
132 :
133 : // Higher priorities are used for user-facing tasks, so that a long backlog of housekeeping work (e.g. reconciling on startup, rescheduling
134 : // things on node changes) does not starve user-facing tasks.
135 : #[derive(Debug, Copy, Clone)]
136 : pub(crate) enum ReconcilerPriority {
137 : Normal,
138 : High,
139 : }
140 :
141 : #[derive(Debug, Copy, Clone)]
142 : pub(crate) struct ReconcilerConfig {
143 : pub(crate) priority: ReconcilerPriority,
144 :
145 : // During live migration give up on warming-up the secondary
146 : // after this timeout.
147 : secondary_warmup_timeout: Option<Duration>,
148 :
149 : // During live migrations this is the amount of time that
150 : // the pagserver will hold our poll.
151 : secondary_download_request_timeout: Option<Duration>,
152 :
153 : // A hint indicating whether this reconciliation is done on the
154 : // creation of a new tenant. This only informs logging behaviour.
155 : tenant_creation_hint: bool,
156 : }
157 :
158 : impl ReconcilerConfig {
159 : /// Configs are always constructed with an explicit priority, to force callers to think about whether
160 : /// the operation they're scheduling is high-priority or not. Normal priority is not a safe default, because
161 : /// scheduling something user-facing at normal priority can result in it getting starved out by background work.
162 0 : pub(crate) fn new(priority: ReconcilerPriority) -> Self {
163 0 : Self {
164 0 : priority,
165 0 : secondary_warmup_timeout: None,
166 0 : secondary_download_request_timeout: None,
167 0 : tenant_creation_hint: false,
168 0 : }
169 0 : }
170 :
171 0 : pub(crate) fn get_secondary_warmup_timeout(&self) -> Duration {
172 : const SECONDARY_WARMUP_TIMEOUT_DEFAULT: Duration = Duration::from_secs(300);
173 0 : self.secondary_warmup_timeout
174 0 : .unwrap_or(SECONDARY_WARMUP_TIMEOUT_DEFAULT)
175 0 : }
176 :
177 0 : pub(crate) fn get_secondary_download_request_timeout(&self) -> Duration {
178 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT_DEFAULT: Duration = Duration::from_secs(20);
179 0 : self.secondary_download_request_timeout
180 0 : .unwrap_or(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT_DEFAULT)
181 0 : }
182 :
183 0 : pub(crate) fn tenant_creation_hint(&self) -> bool {
184 0 : self.tenant_creation_hint
185 0 : }
186 : }
187 :
188 : impl From<&MigrationConfig> for ReconcilerConfig {
189 0 : fn from(value: &MigrationConfig) -> Self {
190 0 : // Run reconciler at high priority because MigrationConfig comes from human requests that should
191 0 : // be presumed urgent.
192 0 : let mut builder = ReconcilerConfigBuilder::new(ReconcilerPriority::High);
193 :
194 0 : if let Some(timeout) = value.secondary_warmup_timeout {
195 0 : builder = builder.secondary_warmup_timeout(timeout)
196 0 : }
197 :
198 0 : if let Some(timeout) = value.secondary_download_request_timeout {
199 0 : builder = builder.secondary_download_request_timeout(timeout)
200 0 : }
201 :
202 0 : builder.build()
203 0 : }
204 : }
205 :
206 : /// RAII resource units granted to a Reconciler, which it should keep alive until it finishes doing I/O
207 : pub(crate) struct ReconcileUnits {
208 : _sem_units: tokio::sync::OwnedSemaphorePermit,
209 : }
210 :
211 : impl ReconcileUnits {
212 0 : pub(crate) fn new(sem_units: tokio::sync::OwnedSemaphorePermit) -> Self {
213 0 : Self {
214 0 : _sem_units: sem_units,
215 0 : }
216 0 : }
217 : }
218 :
219 : /// This is a snapshot of [`crate::tenant_shard::IntentState`], but it does not do any
220 : /// reference counting for Scheduler. The IntentState is what the scheduler works with,
221 : /// and the TargetState is just the instruction for a particular Reconciler run.
222 : #[derive(Debug)]
223 : pub(crate) struct TargetState {
224 : pub(crate) attached: Option<Node>,
225 : pub(crate) secondary: Vec<Node>,
226 : }
227 :
228 : impl TargetState {
229 0 : pub(crate) fn from_intent(nodes: &HashMap<NodeId, Node>, intent: &IntentState) -> Self {
230 0 : Self {
231 0 : attached: intent.get_attached().map(|n| {
232 0 : nodes
233 0 : .get(&n)
234 0 : .expect("Intent attached referenced non-existent node")
235 0 : .clone()
236 0 : }),
237 0 : secondary: intent
238 0 : .get_secondary()
239 0 : .iter()
240 0 : .map(|n| {
241 0 : nodes
242 0 : .get(n)
243 0 : .expect("Intent secondary referenced non-existent node")
244 0 : .clone()
245 0 : })
246 0 : .collect(),
247 0 : }
248 0 : }
249 : }
250 :
251 : #[derive(thiserror::Error, Debug)]
252 : pub(crate) enum ReconcileError {
253 : #[error(transparent)]
254 : Remote(#[from] mgmt_api::Error),
255 : #[error(transparent)]
256 : Notify(#[from] NotifyError),
257 : #[error("Cancelled")]
258 : Cancel,
259 : #[error(transparent)]
260 : Other(#[from] anyhow::Error),
261 : }
262 :
263 : impl Reconciler {
264 0 : async fn location_config(
265 0 : &mut self,
266 0 : node: &Node,
267 0 : config: LocationConfig,
268 0 : flush_ms: Option<Duration>,
269 0 : lazy: bool,
270 0 : ) -> Result<(), ReconcileError> {
271 0 : if !node.is_available() && config.mode == LocationConfigMode::Detached {
272 : // [`crate::service::Service::node_activate_reconcile`] will update the observed state
273 : // when the node comes back online. At that point, the intent and observed states will
274 : // be mismatched and a background reconciliation will detach.
275 0 : tracing::info!(
276 0 : "Node {node} is unavailable during detach: proceeding anyway, it will be detached via background reconciliation"
277 : );
278 0 : return Ok(());
279 0 : }
280 0 :
281 0 : self.observed
282 0 : .locations
283 0 : .insert(node.get_id(), ObservedStateLocation { conf: None });
284 0 :
285 0 : // TODO: amend locations that use long-polling: they will hit this timeout.
286 0 : let timeout = Duration::from_secs(25);
287 0 :
288 0 : tracing::info!("location_config({node}) calling: {:?}", config);
289 0 : let tenant_shard_id = self.tenant_shard_id;
290 0 : let config_ref = &config;
291 0 : match node
292 0 : .with_client_retries(
293 0 : |client| async move {
294 0 : let config = config_ref.clone();
295 0 : client
296 0 : .location_config(tenant_shard_id, config.clone(), flush_ms, lazy)
297 0 : .await
298 0 : },
299 0 : &self.service_config.jwt_token,
300 0 : 1,
301 0 : 3,
302 0 : timeout,
303 0 : &self.cancel,
304 0 : )
305 0 : .await
306 : {
307 0 : Some(Ok(_)) => {}
308 0 : Some(Err(e)) => return Err(e.into()),
309 0 : None => return Err(ReconcileError::Cancel),
310 : };
311 0 : tracing::info!("location_config({node}) complete: {:?}", config);
312 :
313 0 : match config.mode {
314 0 : LocationConfigMode::Detached => {
315 0 : self.observed.locations.remove(&node.get_id());
316 0 : }
317 0 : _ => {
318 0 : self.observed
319 0 : .locations
320 0 : .insert(node.get_id(), ObservedStateLocation { conf: Some(config) });
321 0 : }
322 : }
323 :
324 0 : Ok(())
325 0 : }
326 :
327 0 : fn get_node(&self, node_id: &NodeId) -> Option<&Node> {
328 0 : if let Some(node) = self.intent.attached.as_ref() {
329 0 : if node.get_id() == *node_id {
330 0 : return Some(node);
331 0 : }
332 0 : }
333 :
334 0 : if let Some(node) = self
335 0 : .intent
336 0 : .secondary
337 0 : .iter()
338 0 : .find(|n| n.get_id() == *node_id)
339 : {
340 0 : return Some(node);
341 0 : }
342 :
343 0 : if let Some(node) = self.detach.iter().find(|n| n.get_id() == *node_id) {
344 0 : return Some(node);
345 0 : }
346 0 :
347 0 : None
348 0 : }
349 :
350 0 : async fn maybe_live_migrate(&mut self) -> Result<(), ReconcileError> {
351 0 : let destination = if let Some(node) = &self.intent.attached {
352 0 : match self.observed.locations.get(&node.get_id()) {
353 0 : Some(conf) => {
354 : // We will do a live migration only if the intended destination is not
355 : // currently in an attached state.
356 0 : match &conf.conf {
357 0 : Some(conf) if conf.mode == LocationConfigMode::Secondary => {
358 0 : // Fall through to do a live migration
359 0 : node
360 : }
361 : None | Some(_) => {
362 : // Attached or uncertain: don't do a live migration, proceed
363 : // with a general-case reconciliation
364 0 : tracing::info!("maybe_live_migrate: destination is None or attached");
365 0 : return Ok(());
366 : }
367 : }
368 : }
369 : None => {
370 : // Our destination is not attached: maybe live migrate if some other
371 : // node is currently attached. Fall through.
372 0 : node
373 : }
374 : }
375 : } else {
376 : // No intent to be attached
377 0 : tracing::info!("maybe_live_migrate: no attached intent");
378 0 : return Ok(());
379 : };
380 :
381 0 : let mut origin = None;
382 0 : for (node_id, state) in &self.observed.locations {
383 0 : if let Some(observed_conf) = &state.conf {
384 0 : if observed_conf.mode == LocationConfigMode::AttachedSingle {
385 : // We will only attempt live migration if the origin is not offline: this
386 : // avoids trying to do it while reconciling after responding to an HA failover.
387 0 : if let Some(node) = self.get_node(node_id) {
388 0 : if node.is_available() {
389 0 : origin = Some(node.clone());
390 0 : break;
391 0 : }
392 0 : }
393 0 : }
394 0 : }
395 : }
396 :
397 0 : let Some(origin) = origin else {
398 0 : tracing::info!("maybe_live_migrate: no origin found");
399 0 : return Ok(());
400 : };
401 :
402 : // We have an origin and a destination: proceed to do the live migration
403 0 : tracing::info!("Live migrating {}->{}", origin, destination);
404 0 : self.live_migrate(origin, destination.clone()).await?;
405 :
406 0 : Ok(())
407 0 : }
408 :
409 0 : async fn wait_lsn(
410 0 : &self,
411 0 : node: &Node,
412 0 : tenant_shard_id: TenantShardId,
413 0 : timelines: HashMap<TimelineId, Lsn>,
414 0 : ) -> Result<StatusCode, ReconcileError> {
415 : const TIMEOUT: Duration = Duration::from_secs(10);
416 :
417 0 : let client = PageserverClient::new(
418 0 : node.get_id(),
419 0 : node.base_url(),
420 0 : self.service_config.jwt_token.as_deref(),
421 0 : );
422 0 :
423 0 : client
424 0 : .wait_lsn(
425 0 : tenant_shard_id,
426 0 : TenantWaitLsnRequest {
427 0 : timelines,
428 0 : timeout: TIMEOUT,
429 0 : },
430 0 : )
431 0 : .await
432 0 : .map_err(|e| e.into())
433 0 : }
434 :
435 0 : async fn get_lsns(
436 0 : &self,
437 0 : tenant_shard_id: TenantShardId,
438 0 : node: &Node,
439 0 : ) -> anyhow::Result<HashMap<TimelineId, Lsn>> {
440 0 : let client = PageserverClient::new(
441 0 : node.get_id(),
442 0 : node.base_url(),
443 0 : self.service_config.jwt_token.as_deref(),
444 0 : );
445 :
446 0 : let timelines = client.timeline_list(&tenant_shard_id).await?;
447 0 : Ok(timelines
448 0 : .into_iter()
449 0 : .map(|t| (t.timeline_id, t.last_record_lsn))
450 0 : .collect())
451 0 : }
452 :
453 0 : async fn secondary_download(
454 0 : &self,
455 0 : tenant_shard_id: TenantShardId,
456 0 : node: &Node,
457 0 : ) -> Result<(), ReconcileError> {
458 0 : // This is not the timeout for a request, but the total amount of time we're willing to wait
459 0 : // for a secondary location to get up to date before
460 0 : let total_download_timeout = self.reconciler_config.get_secondary_warmup_timeout();
461 0 :
462 0 : // This the long-polling interval for the secondary download requests we send to destination pageserver
463 0 : // during a migration.
464 0 : let request_download_timeout = self
465 0 : .reconciler_config
466 0 : .get_secondary_download_request_timeout();
467 0 :
468 0 : let started_at = Instant::now();
469 :
470 : loop {
471 0 : let (status, progress) = match node
472 0 : .with_client_retries(
473 0 : |client| async move {
474 0 : client
475 0 : .tenant_secondary_download(
476 0 : tenant_shard_id,
477 0 : Some(request_download_timeout),
478 0 : )
479 0 : .await
480 0 : },
481 0 : &self.service_config.jwt_token,
482 0 : 1,
483 0 : 3,
484 0 : request_download_timeout * 2,
485 0 : &self.cancel,
486 0 : )
487 0 : .await
488 : {
489 0 : None => Err(ReconcileError::Cancel),
490 0 : Some(Ok(v)) => Ok(v),
491 0 : Some(Err(e)) => {
492 0 : // Give up, but proceed: it's unfortunate if we couldn't freshen the destination before
493 0 : // attaching, but we should not let an issue with a secondary location stop us proceeding
494 0 : // with a live migration.
495 0 : tracing::warn!("Failed to prepare by downloading layers on node {node}: {e})");
496 0 : return Ok(());
497 : }
498 0 : }?;
499 :
500 0 : if status == StatusCode::OK {
501 0 : tracing::info!(
502 0 : "Downloads to {} complete: {}/{} layers, {}/{} bytes",
503 : node,
504 : progress.layers_downloaded,
505 : progress.layers_total,
506 : progress.bytes_downloaded,
507 : progress.bytes_total
508 : );
509 0 : return Ok(());
510 0 : } else if status == StatusCode::ACCEPTED {
511 0 : let total_runtime = started_at.elapsed();
512 0 : if total_runtime > total_download_timeout {
513 0 : tracing::warn!("Timed out after {}ms downloading layers to {node}. Progress so far: {}/{} layers, {}/{} bytes",
514 0 : total_runtime.as_millis(),
515 : progress.layers_downloaded,
516 : progress.layers_total,
517 : progress.bytes_downloaded,
518 : progress.bytes_total
519 : );
520 : // Give up, but proceed: an incompletely warmed destination doesn't prevent migration working,
521 : // it just makes the I/O performance for users less good.
522 0 : return Ok(());
523 0 : }
524 0 :
525 0 : // Log and proceed around the loop to retry. We don't sleep between requests, because our HTTP call
526 0 : // to the pageserver is a long-poll.
527 0 : tracing::info!(
528 0 : "Downloads to {} not yet complete: {}/{} layers, {}/{} bytes",
529 : node,
530 : progress.layers_downloaded,
531 : progress.layers_total,
532 : progress.bytes_downloaded,
533 : progress.bytes_total
534 : );
535 0 : }
536 : }
537 0 : }
538 :
539 : /// This function does _not_ mutate any state, so it is cancellation safe.
540 : ///
541 : /// This function does not respect [`Self::cancel`], callers should handle that.
542 0 : async fn await_lsn(
543 0 : &self,
544 0 : tenant_shard_id: TenantShardId,
545 0 : node: &Node,
546 0 : baseline: HashMap<TimelineId, Lsn>,
547 0 : ) -> anyhow::Result<()> {
548 : // Signal to the pageserver that it should ingest up to the baseline LSNs.
549 : loop {
550 0 : match self.wait_lsn(node, tenant_shard_id, baseline.clone()).await {
551 : Ok(StatusCode::OK) => {
552 : // Everything is caught up
553 0 : return Ok(());
554 : }
555 : Ok(StatusCode::ACCEPTED) => {
556 : // Some timelines are not caught up yet.
557 : // They'll be polled below.
558 0 : break;
559 : }
560 : Ok(StatusCode::NOT_FOUND) => {
561 : // None of the timelines are present on the pageserver.
562 : // This is correct if they've all been deleted, but
563 : // let let the polling loop below cross check.
564 0 : break;
565 : }
566 0 : Ok(status_code) => {
567 0 : tracing::warn!(
568 0 : "Unexpected status code ({status_code}) returned by wait_lsn endpoint"
569 : );
570 0 : break;
571 : }
572 0 : Err(e) => {
573 0 : tracing::info!("🕑 Can't trigger LSN wait on {node} yet, waiting ({e})",);
574 0 : tokio::time::sleep(Duration::from_millis(500)).await;
575 0 : continue;
576 : }
577 : }
578 : }
579 :
580 : // Poll the LSNs until they catch up
581 : loop {
582 0 : let latest = match self.get_lsns(tenant_shard_id, node).await {
583 0 : Ok(l) => l,
584 0 : Err(e) => {
585 0 : tracing::info!("🕑 Can't get LSNs on node {node} yet, waiting ({e})",);
586 0 : tokio::time::sleep(Duration::from_millis(500)).await;
587 0 : continue;
588 : }
589 : };
590 :
591 0 : let mut any_behind: bool = false;
592 0 : for (timeline_id, baseline_lsn) in &baseline {
593 0 : match latest.get(timeline_id) {
594 0 : Some(latest_lsn) => {
595 0 : tracing::info!(timeline_id = %timeline_id, "🕑 LSN origin {baseline_lsn} vs destination {latest_lsn}");
596 0 : if latest_lsn < baseline_lsn {
597 0 : any_behind = true;
598 0 : }
599 : }
600 0 : None => {
601 0 : // Timeline was deleted in the meantime - ignore it
602 0 : }
603 : }
604 : }
605 :
606 0 : if !any_behind {
607 0 : tracing::info!("✅ LSN caught up. Proceeding...");
608 0 : break;
609 : } else {
610 0 : tokio::time::sleep(Duration::from_millis(500)).await;
611 : }
612 : }
613 :
614 0 : Ok(())
615 0 : }
616 :
617 0 : pub async fn live_migrate(
618 0 : &mut self,
619 0 : origin_ps: Node,
620 0 : dest_ps: Node,
621 0 : ) -> Result<(), ReconcileError> {
622 0 : // `maybe_live_migrate` is responsibble for sanity of inputs
623 0 : assert!(origin_ps.get_id() != dest_ps.get_id());
624 :
625 0 : fn build_location_config(
626 0 : shard: &ShardIdentity,
627 0 : config: &TenantConfig,
628 0 : mode: LocationConfigMode,
629 0 : generation: Option<Generation>,
630 0 : secondary_conf: Option<LocationConfigSecondary>,
631 0 : ) -> LocationConfig {
632 0 : LocationConfig {
633 0 : mode,
634 0 : generation: generation.map(|g| g.into().unwrap()),
635 0 : secondary_conf,
636 0 : tenant_conf: config.clone(),
637 0 : shard_number: shard.number.0,
638 0 : shard_count: shard.count.literal(),
639 0 : shard_stripe_size: shard.stripe_size.0,
640 0 : }
641 0 : }
642 :
643 0 : tracing::info!("🔁 Switching origin node {origin_ps} to stale mode",);
644 :
645 : // FIXME: it is incorrect to use self.generation here, we should use the generation
646 : // from the ObservedState of the origin pageserver (it might be older than self.generation)
647 0 : let stale_conf = build_location_config(
648 0 : &self.shard,
649 0 : &self.config,
650 0 : LocationConfigMode::AttachedStale,
651 0 : self.generation,
652 0 : None,
653 0 : );
654 0 : self.location_config(&origin_ps, stale_conf, Some(Duration::from_secs(10)), false)
655 0 : .await?;
656 :
657 0 : let baseline_lsns = Some(self.get_lsns(self.tenant_shard_id, &origin_ps).await?);
658 :
659 : // If we are migrating to a destination that has a secondary location, warm it up first
660 0 : if let Some(destination_conf) = self.observed.locations.get(&dest_ps.get_id()) {
661 0 : if let Some(destination_conf) = &destination_conf.conf {
662 0 : if destination_conf.mode == LocationConfigMode::Secondary {
663 0 : tracing::info!("🔁 Downloading latest layers to destination node {dest_ps}",);
664 0 : self.secondary_download(self.tenant_shard_id, &dest_ps)
665 0 : .await?;
666 0 : }
667 0 : }
668 0 : }
669 :
670 0 : pausable_failpoint!("reconciler-live-migrate-pre-generation-inc");
671 :
672 : // Increment generation before attaching to new pageserver
673 : self.generation = Some(
674 0 : self.persistence
675 0 : .increment_generation(self.tenant_shard_id, dest_ps.get_id())
676 0 : .await?,
677 : );
678 :
679 0 : let dest_conf = build_location_config(
680 0 : &self.shard,
681 0 : &self.config,
682 0 : LocationConfigMode::AttachedMulti,
683 0 : self.generation,
684 0 : None,
685 0 : );
686 0 :
687 0 : tracing::info!("🔁 Attaching to pageserver {dest_ps}");
688 0 : self.location_config(&dest_ps, dest_conf, None, false)
689 0 : .await?;
690 :
691 0 : pausable_failpoint!("reconciler-live-migrate-pre-await-lsn");
692 :
693 0 : if let Some(baseline) = baseline_lsns {
694 0 : tracing::info!("🕑 Waiting for LSN to catch up...");
695 0 : tokio::select! {
696 0 : r = self.await_lsn(self.tenant_shard_id, &dest_ps, baseline) => {r?;}
697 0 : _ = self.cancel.cancelled() => {return Err(ReconcileError::Cancel)}
698 : };
699 0 : }
700 :
701 0 : tracing::info!("🔁 Notifying compute to use pageserver {dest_ps}");
702 :
703 : // During a live migration it is unhelpful to proceed if we couldn't notify compute: if we detach
704 : // the origin without notifying compute, we will render the tenant unavailable.
705 0 : self.compute_notify_blocking(&origin_ps).await?;
706 0 : pausable_failpoint!("reconciler-live-migrate-post-notify");
707 :
708 : // Downgrade the origin to secondary. If the tenant's policy is PlacementPolicy::Attached(0), then
709 : // this location will be deleted in the general case reconciliation that runs after this.
710 0 : let origin_secondary_conf = build_location_config(
711 0 : &self.shard,
712 0 : &self.config,
713 0 : LocationConfigMode::Secondary,
714 0 : None,
715 0 : Some(LocationConfigSecondary { warm: true }),
716 0 : );
717 0 : self.location_config(&origin_ps, origin_secondary_conf.clone(), None, false)
718 0 : .await?;
719 : // TODO: we should also be setting the ObservedState on earlier API calls, in case we fail
720 : // partway through. In fact, all location conf API calls should be in a wrapper that sets
721 : // the observed state to None, then runs, then sets it to what we wrote.
722 0 : self.observed.locations.insert(
723 0 : origin_ps.get_id(),
724 0 : ObservedStateLocation {
725 0 : conf: Some(origin_secondary_conf),
726 0 : },
727 0 : );
728 0 :
729 0 : pausable_failpoint!("reconciler-live-migrate-post-detach");
730 :
731 0 : tracing::info!("🔁 Switching to AttachedSingle mode on node {dest_ps}",);
732 0 : let dest_final_conf = build_location_config(
733 0 : &self.shard,
734 0 : &self.config,
735 0 : LocationConfigMode::AttachedSingle,
736 0 : self.generation,
737 0 : None,
738 0 : );
739 0 : self.location_config(&dest_ps, dest_final_conf.clone(), None, false)
740 0 : .await?;
741 0 : self.observed.locations.insert(
742 0 : dest_ps.get_id(),
743 0 : ObservedStateLocation {
744 0 : conf: Some(dest_final_conf),
745 0 : },
746 0 : );
747 0 :
748 0 : tracing::info!("✅ Migration complete");
749 :
750 0 : Ok(())
751 0 : }
752 :
753 0 : async fn maybe_refresh_observed(&mut self) -> Result<(), ReconcileError> {
754 : // If the attached node has uncertain state, read it from the pageserver before proceeding: this
755 : // is important to avoid spurious generation increments.
756 : //
757 : // We don't need to do this for secondary/detach locations because it's harmless to just PUT their
758 : // location conf, whereas for attached locations it can interrupt clients if we spuriously destroy/recreate
759 : // the `Timeline` object in the pageserver.
760 :
761 0 : let Some(attached_node) = self.intent.attached.as_ref() else {
762 : // Nothing to do
763 0 : return Ok(());
764 : };
765 :
766 0 : if matches!(
767 0 : self.observed.locations.get(&attached_node.get_id()),
768 : Some(ObservedStateLocation { conf: None })
769 : ) {
770 0 : let tenant_shard_id = self.tenant_shard_id;
771 0 : let observed_conf = match attached_node
772 0 : .with_client_retries(
773 0 : |client| async move { client.get_location_config(tenant_shard_id).await },
774 0 : &self.service_config.jwt_token,
775 0 : 1,
776 0 : 1,
777 0 : Duration::from_secs(5),
778 0 : &self.cancel,
779 0 : )
780 0 : .await
781 : {
782 0 : Some(Ok(observed)) => Some(observed),
783 0 : Some(Err(mgmt_api::Error::ApiError(status, _msg)))
784 0 : if status == StatusCode::NOT_FOUND =>
785 0 : {
786 0 : None
787 : }
788 0 : Some(Err(e)) => return Err(e.into()),
789 0 : None => return Err(ReconcileError::Cancel),
790 : };
791 0 : tracing::info!("Scanned location configuration on {attached_node}: {observed_conf:?}");
792 0 : match observed_conf {
793 0 : Some(conf) => {
794 0 : // Pageserver returned a state: update it in observed. This may still be an indeterminate (None) state,
795 0 : // if internally the pageserver's TenantSlot was being mutated (e.g. some long running API call is still running)
796 0 : self.observed
797 0 : .locations
798 0 : .insert(attached_node.get_id(), ObservedStateLocation { conf });
799 0 : }
800 0 : None => {
801 0 : // Pageserver returned 404: we have confirmation that there is no state for this shard on that pageserver.
802 0 : self.observed.locations.remove(&attached_node.get_id());
803 0 : }
804 : }
805 0 : }
806 :
807 0 : Ok(())
808 0 : }
809 :
810 : /// Reconciling a tenant makes API calls to pageservers until the observed state
811 : /// matches the intended state.
812 : ///
813 : /// First we apply special case handling (e.g. for live migrations), and then a
814 : /// general case reconciliation where we walk through the intent by pageserver
815 : /// and call out to the pageserver to apply the desired state.
816 : ///
817 : /// An Ok(()) result indicates that we successfully attached the tenant, but _not_ that
818 : /// all locations for the tenant are in the expected state. When nodes that are to be detached
819 : /// or configured as secondary are unavailable, we may return Ok(()) but leave the shard in a
820 : /// state where it still requires later reconciliation.
821 0 : pub(crate) async fn reconcile(&mut self) -> Result<(), ReconcileError> {
822 0 : // Prepare: if we have uncertain `observed` state for our would-be attachement location, then refresh it
823 0 : self.maybe_refresh_observed().await?;
824 :
825 : // Special case: live migration
826 0 : self.maybe_live_migrate().await?;
827 :
828 : // If the attached pageserver is not attached, do so now.
829 0 : if let Some(node) = self.intent.attached.as_ref() {
830 : // If we are in an attached policy, then generation must have been set (null generations
831 : // are only present when a tenant is initially loaded with a secondary policy)
832 0 : debug_assert!(self.generation.is_some());
833 0 : let Some(generation) = self.generation else {
834 0 : return Err(ReconcileError::Other(anyhow::anyhow!(
835 0 : "Attempted to attach with NULL generation"
836 0 : )));
837 : };
838 :
839 0 : let mut wanted_conf = attached_location_conf(
840 0 : generation,
841 0 : &self.shard,
842 0 : &self.config,
843 0 : &self.placement_policy,
844 0 : );
845 0 : match self.observed.locations.get(&node.get_id()) {
846 0 : Some(conf) if conf.conf.as_ref() == Some(&wanted_conf) => {
847 0 : // Nothing to do
848 0 : tracing::info!(node_id=%node.get_id(), "Observed configuration already correct.")
849 : }
850 0 : observed => {
851 : // In all cases other than a matching observed configuration, we will
852 : // reconcile this location. This includes locations with different configurations, as well
853 : // as locations with unknown (None) observed state.
854 :
855 : // Incrementing generation is the safe general case, but is inefficient for changes that only
856 : // modify some details (e.g. the tenant's config).
857 0 : let increment_generation = match observed {
858 0 : None => true,
859 0 : Some(ObservedStateLocation { conf: None }) => true,
860 : Some(ObservedStateLocation {
861 0 : conf: Some(observed),
862 0 : }) => {
863 0 : let generations_match = observed.generation == wanted_conf.generation;
864 0 :
865 0 : // We may skip incrementing the generation if the location is already in the expected mode and
866 0 : // generation. In principle it would also be safe to skip from certain other modes (e.g. AttachedStale),
867 0 : // but such states are handled inside `live_migrate`, and if we see that state here we're cleaning up
868 0 : // after a restart/crash, so fall back to the universally safe path of incrementing generation.
869 0 : !generations_match || (observed.mode != wanted_conf.mode)
870 : }
871 : };
872 :
873 0 : if increment_generation {
874 0 : pausable_failpoint!("reconciler-pre-increment-generation");
875 :
876 0 : let generation = self
877 0 : .persistence
878 0 : .increment_generation(self.tenant_shard_id, node.get_id())
879 0 : .await?;
880 0 : self.generation = Some(generation);
881 0 : wanted_conf.generation = generation.into();
882 0 : }
883 0 : tracing::info!(node_id=%node.get_id(), "Observed configuration requires update.");
884 :
885 : // Because `node` comes from a ref to &self, clone it before calling into a &mut self
886 : // function: this could be avoided by refactoring the state mutated by location_config into
887 : // a separate type to Self.
888 0 : let node = node.clone();
889 0 :
890 0 : // Use lazy=true, because we may run many of Self concurrently, and do not want to
891 0 : // overload the pageserver with logical size calculations.
892 0 : self.location_config(&node, wanted_conf, None, true).await?;
893 0 : self.compute_notify().await?;
894 : }
895 : }
896 0 : }
897 :
898 : // Configure secondary locations: if these were previously attached this
899 : // implicitly downgrades them from attached to secondary.
900 0 : let mut changes = Vec::new();
901 0 : for node in &self.intent.secondary {
902 0 : let wanted_conf = secondary_location_conf(&self.shard, &self.config);
903 0 : match self.observed.locations.get(&node.get_id()) {
904 0 : Some(conf) if conf.conf.as_ref() == Some(&wanted_conf) => {
905 0 : // Nothing to do
906 0 : tracing::info!(node_id=%node.get_id(), "Observed configuration already correct.")
907 : }
908 : _ => {
909 : // Only try and configure secondary locations on nodes that are available. This
910 : // allows the reconciler to "succeed" while some secondaries are offline (e.g. after
911 : // a node failure, where the failed node will have a secondary intent)
912 0 : if node.is_available() {
913 0 : tracing::info!(node_id=%node.get_id(), "Observed configuration requires update.");
914 0 : changes.push((node.clone(), wanted_conf))
915 : } else {
916 0 : tracing::info!(node_id=%node.get_id(), "Skipping configuration as secondary, node is unavailable");
917 0 : self.observed
918 0 : .locations
919 0 : .insert(node.get_id(), ObservedStateLocation { conf: None });
920 : }
921 : }
922 : }
923 : }
924 :
925 : // Detach any extraneous pageservers that are no longer referenced
926 : // by our intent.
927 0 : for node in &self.detach {
928 0 : changes.push((
929 0 : node.clone(),
930 0 : LocationConfig {
931 0 : mode: LocationConfigMode::Detached,
932 0 : generation: None,
933 0 : secondary_conf: None,
934 0 : shard_number: self.shard.number.0,
935 0 : shard_count: self.shard.count.literal(),
936 0 : shard_stripe_size: self.shard.stripe_size.0,
937 0 : tenant_conf: self.config.clone(),
938 0 : },
939 0 : ));
940 0 : }
941 :
942 0 : for (node, conf) in changes {
943 0 : if self.cancel.is_cancelled() {
944 0 : return Err(ReconcileError::Cancel);
945 0 : }
946 0 : // We only try to configure secondary locations if the node is available. This does
947 0 : // not stop us succeeding with the reconcile, because our core goal is to make the
948 0 : // shard _available_ (the attached location), and configuring secondary locations
949 0 : // can be done lazily when the node becomes available (via background reconciliation).
950 0 : if node.is_available() {
951 0 : self.location_config(&node, conf, None, false).await?;
952 : } else {
953 : // If the node is unavailable, we skip and consider the reconciliation successful: this
954 : // is a common case where a pageserver is marked unavailable: we demote a location on
955 : // that unavailable pageserver to secondary.
956 0 : tracing::info!("Skipping configuring secondary location {node}, it is unavailable");
957 0 : self.observed
958 0 : .locations
959 0 : .insert(node.get_id(), ObservedStateLocation { conf: None });
960 : }
961 : }
962 :
963 : // The condition below identifies a detach. We must have no attached intent and
964 : // must have been attached to something previously. Pass this information to
965 : // the [`ComputeHook`] such that it can update its tenant-wide state.
966 0 : if self.intent.attached.is_none() && !self.detach.is_empty() {
967 0 : // TODO: Consider notifying control plane about detaches. This would avoid situations
968 0 : // where the compute tries to start-up with a stale set of pageservers.
969 0 : self.compute_hook
970 0 : .handle_detach(self.tenant_shard_id, self.shard.stripe_size);
971 0 : }
972 :
973 0 : pausable_failpoint!("reconciler-epilogue");
974 :
975 0 : Ok(())
976 0 : }
977 :
978 0 : pub(crate) async fn compute_notify(&mut self) -> Result<(), NotifyError> {
979 : // Whenever a particular Reconciler emits a notification, it is always notifying for the intended
980 : // destination.
981 0 : if let Some(node) = &self.intent.attached {
982 0 : let result = self
983 0 : .compute_hook
984 0 : .notify(
985 0 : compute_hook::ShardUpdate {
986 0 : tenant_shard_id: self.tenant_shard_id,
987 0 : node_id: node.get_id(),
988 0 : stripe_size: self.shard.stripe_size,
989 0 : preferred_az: self.preferred_az.as_ref().map(Cow::Borrowed),
990 0 : },
991 0 : &self.cancel,
992 0 : )
993 0 : .await;
994 0 : if let Err(e) = &result {
995 : // Set this flag so that in our ReconcileResult we will set the flag on the shard that it
996 : // needs to retry at some point.
997 0 : self.compute_notify_failure = true;
998 :
999 : // It is up to the caller whether they want to drop out on this error, but they don't have to:
1000 : // in general we should avoid letting unavailability of the cloud control plane stop us from
1001 : // making progress.
1002 0 : match e {
1003 0 : // 404s from cplane during tenant creation are expected.
1004 0 : // Cplane only persists the shards to the database after
1005 0 : // creating the tenant and the timeline. If we notify before
1006 0 : // that, we'll get a 404.
1007 0 : //
1008 0 : // This is fine because tenant creations happen via /location_config
1009 0 : // and that returns the list of locations in the response. Hence, we
1010 0 : // silence the error and return Ok(()) here. Reconciliation will still
1011 0 : // be retried because we set [`Reconciler::compute_notify_failure`] above.
1012 0 : NotifyError::Unexpected(hyper::StatusCode::NOT_FOUND)
1013 0 : if self.reconciler_config.tenant_creation_hint() =>
1014 0 : {
1015 0 : return Ok(());
1016 : }
1017 0 : NotifyError::ShuttingDown => {}
1018 : _ => {
1019 0 : tracing::warn!(
1020 0 : "Failed to notify compute of attached pageserver {node}: {e}"
1021 : );
1022 : }
1023 : }
1024 0 : }
1025 0 : result
1026 : } else {
1027 0 : Ok(())
1028 : }
1029 0 : }
1030 :
1031 : /// Compare the observed state snapshot from when the reconcile was created
1032 : /// with the final observed state in order to generate observed state deltas.
1033 0 : pub(crate) fn observed_deltas(&self) -> Vec<ObservedStateDelta> {
1034 0 : let mut deltas = Vec::default();
1035 :
1036 0 : for (node_id, location) in &self.observed.locations {
1037 0 : let previous_location = self.original_observed.locations.get(node_id);
1038 0 : let do_upsert = match previous_location {
1039 : // Location config changed for node
1040 0 : Some(prev) if location.conf != prev.conf => true,
1041 : // New location config for node
1042 0 : None => true,
1043 : // Location config has not changed for node
1044 0 : _ => false,
1045 : };
1046 :
1047 0 : if do_upsert {
1048 0 : deltas.push(ObservedStateDelta::Upsert(Box::new((
1049 0 : *node_id,
1050 0 : location.clone(),
1051 0 : ))));
1052 0 : }
1053 : }
1054 :
1055 0 : for node_id in self.original_observed.locations.keys() {
1056 0 : if !self.observed.locations.contains_key(node_id) {
1057 0 : deltas.push(ObservedStateDelta::Delete(*node_id));
1058 0 : }
1059 : }
1060 :
1061 0 : deltas
1062 0 : }
1063 :
1064 : /// Keep trying to notify the compute indefinitely, only dropping out if:
1065 : /// - the node `origin` becomes unavailable -> Ok(())
1066 : /// - the node `origin` no longer has our tenant shard attached -> Ok(())
1067 : /// - our cancellation token fires -> Err(ReconcileError::Cancelled)
1068 : ///
1069 : /// This is used during live migration, where we do not wish to detach
1070 : /// an origin location until the compute definitely knows about the new
1071 : /// location.
1072 : ///
1073 : /// In cases where the origin node becomes unavailable, we return success, indicating
1074 : /// to the caller that they should continue irrespective of whether the compute was notified,
1075 : /// because the origin node is unusable anyway. Notification will be retried later via the
1076 : /// [`Self::compute_notify_failure`] flag.
1077 0 : async fn compute_notify_blocking(&mut self, origin: &Node) -> Result<(), ReconcileError> {
1078 0 : let mut notify_attempts = 0;
1079 0 : while let Err(e) = self.compute_notify().await {
1080 0 : match e {
1081 0 : NotifyError::Fatal(_) => return Err(ReconcileError::Notify(e)),
1082 0 : NotifyError::ShuttingDown => return Err(ReconcileError::Cancel),
1083 : _ => {
1084 0 : tracing::warn!(
1085 0 : "Live migration blocked by compute notification error, retrying: {e}"
1086 : );
1087 : }
1088 : }
1089 :
1090 : // Did the origin pageserver become unavailable?
1091 0 : if !origin.is_available() {
1092 0 : tracing::info!("Giving up on compute notification because {origin} is unavailable");
1093 0 : break;
1094 0 : }
1095 0 :
1096 0 : // Does the origin pageserver still host the shard we are interested in? We should only
1097 0 : // continue waiting for compute notification to be acked if the old location is still usable.
1098 0 : let tenant_shard_id = self.tenant_shard_id;
1099 0 : match origin
1100 0 : .with_client_retries(
1101 0 : |client| async move { client.get_location_config(tenant_shard_id).await },
1102 0 : &self.service_config.jwt_token,
1103 0 : 1,
1104 0 : 3,
1105 0 : Duration::from_secs(5),
1106 0 : &self.cancel,
1107 0 : )
1108 0 : .await
1109 : {
1110 0 : Some(Ok(Some(location_conf))) => {
1111 0 : if matches!(
1112 0 : location_conf.mode,
1113 : LocationConfigMode::AttachedMulti
1114 : | LocationConfigMode::AttachedSingle
1115 : | LocationConfigMode::AttachedStale
1116 : ) {
1117 0 : tracing::debug!(
1118 0 : "Still attached to {origin}, will wait & retry compute notification"
1119 : );
1120 : } else {
1121 0 : tracing::info!(
1122 0 : "Giving up on compute notification because {origin} is in state {:?}",
1123 : location_conf.mode
1124 : );
1125 0 : return Ok(());
1126 : }
1127 : // Fall through
1128 : }
1129 : Some(Ok(None)) => {
1130 0 : tracing::info!(
1131 0 : "No longer attached to {origin}, giving up on compute notification"
1132 : );
1133 0 : return Ok(());
1134 : }
1135 0 : Some(Err(e)) => {
1136 0 : match e {
1137 : mgmt_api::Error::Cancelled => {
1138 0 : tracing::info!(
1139 0 : "Giving up on compute notification because {origin} is unavailable"
1140 : );
1141 0 : return Ok(());
1142 : }
1143 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, _) => {
1144 0 : tracing::info!(
1145 0 : "No longer attached to {origin}, giving up on compute notification"
1146 : );
1147 0 : return Ok(());
1148 : }
1149 0 : e => {
1150 0 : // Other API errors are unexpected here.
1151 0 : tracing::warn!("Unexpected error checking location on {origin}: {e}");
1152 :
1153 : // Fall through, we will retry compute notification.
1154 : }
1155 : }
1156 : }
1157 0 : None => return Err(ReconcileError::Cancel),
1158 : };
1159 :
1160 0 : exponential_backoff(
1161 0 : notify_attempts,
1162 0 : // Generous waits: control plane operations which might be blocking us usually complete on the order
1163 0 : // of hundreds to thousands of milliseconds, so no point busy polling.
1164 0 : 1.0,
1165 0 : 10.0,
1166 0 : &self.cancel,
1167 0 : )
1168 0 : .await;
1169 0 : notify_attempts += 1;
1170 : }
1171 :
1172 0 : Ok(())
1173 0 : }
1174 : }
1175 :
1176 : /// We tweak the externally-set TenantConfig while configuring
1177 : /// locations, using our awareness of whether secondary locations
1178 : /// are in use to automatically enable/disable heatmap uploads.
1179 0 : fn ha_aware_config(config: &TenantConfig, has_secondaries: bool) -> TenantConfig {
1180 0 : let mut config = config.clone();
1181 0 : if has_secondaries {
1182 0 : if config.heatmap_period.is_none() {
1183 0 : config.heatmap_period = Some(DEFAULT_HEATMAP_PERIOD.to_string());
1184 0 : }
1185 0 : } else {
1186 0 : config.heatmap_period = None;
1187 0 : }
1188 0 : config
1189 0 : }
1190 :
1191 0 : pub(crate) fn attached_location_conf(
1192 0 : generation: Generation,
1193 0 : shard: &ShardIdentity,
1194 0 : config: &TenantConfig,
1195 0 : policy: &PlacementPolicy,
1196 0 : ) -> LocationConfig {
1197 0 : let has_secondaries = match policy {
1198 : PlacementPolicy::Attached(0) | PlacementPolicy::Detached | PlacementPolicy::Secondary => {
1199 0 : false
1200 : }
1201 0 : PlacementPolicy::Attached(_) => true,
1202 : };
1203 :
1204 0 : LocationConfig {
1205 0 : mode: LocationConfigMode::AttachedSingle,
1206 0 : generation: generation.into(),
1207 0 : secondary_conf: None,
1208 0 : shard_number: shard.number.0,
1209 0 : shard_count: shard.count.literal(),
1210 0 : shard_stripe_size: shard.stripe_size.0,
1211 0 : tenant_conf: ha_aware_config(config, has_secondaries),
1212 0 : }
1213 0 : }
1214 :
1215 0 : pub(crate) fn secondary_location_conf(
1216 0 : shard: &ShardIdentity,
1217 0 : config: &TenantConfig,
1218 0 : ) -> LocationConfig {
1219 0 : LocationConfig {
1220 0 : mode: LocationConfigMode::Secondary,
1221 0 : generation: None,
1222 0 : secondary_conf: Some(LocationConfigSecondary { warm: true }),
1223 0 : shard_number: shard.number.0,
1224 0 : shard_count: shard.count.literal(),
1225 0 : shard_stripe_size: shard.stripe_size.0,
1226 0 : tenant_conf: ha_aware_config(config, true),
1227 0 : }
1228 0 : }
|
